# machine-learning
- u
{K EY1} (Kei)
09/19/2022, 2:03 PMNot in the case of every synth - m
mepc36
09/19/2022, 2:03 PMSo would the best model to use be the one with the smallest positive loss_mel? - u
{K EY1} (Kei)
09/19/2022, 2:03 PMIn nnsvs for example, on mdn timelag and duration negative loss is good - u
{K EY1} (Kei)
09/19/2022, 2:03 PMBut on tacotron it would be severely overfit - m
mepc36
09/19/2022, 2:03 PMmmmmmm rough - u
{K EY1} (Kei)
09/19/2022, 2:04 PMPossibly? I don't know much about radtts I'd say either train it for the default number of iterations, or test it as you go along to make sure it's not overfit - h
hecko
09/19/2022, 2:06 PMthe best model is the one that sounds best - h
hecko
09/19/2022, 2:07 PMwith tacotron i sometimes prefer the model with fewer epochs and higher loss - h
hecko
09/19/2022, 2:07 PM(though it seems that everypony disagrees and instead prefers to train for literally thousands of epochs) - u
{K EY1} (Kei)
09/19/2022, 2:08 PMI've trained for hundreds of epochs with tacotron but i don't think i've ever hit the thousands... - u
(Dawn) Will Draw Fictional Women
09/19/2022, 2:13 PMstroke free since 2021 - m
mepc36
09/19/2022, 2:14 PMSo manual testing of synthesized wavs is the best way to pick a model you think? Genuinely interested... - u
{K EY1} (Kei)
09/19/2022, 2:14 PMYes - m
mepc36
09/19/2022, 2:14 PMoooooo very nice, thank you everyone - h
hecko
09/19/2022, 2:15 PMi will note that the differences can be very subtle - m
mepc36
09/19/2022, 3:53 PMyou're right! 2 diff models trained 200 iters apart - the first pronounced the word like 'ballING" but the other pronounced it like "BALLing" - p
PixPrucer
09/19/2022, 4:14 PMYep he's cleaner now - p
PixPrucer
09/19/2022, 4:19 PMNow I just need to obtain the rest of the samples and he's getting released 💥 - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:19 PMi wanna try training on another language but i dont Know any of the ones with good base models - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:19 PMlike i have Ideas and i have Data and i have Raw Coursing Tacotron God Genes i just cannot Form Transcripts - c
Couch
09/19/2022, 4:23 PM#903852974156165170 and #903712295534882897: - c
Couch
09/19/2022, 4:23 PM - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:23 PMtime to Beg on the Internet for other language knowers to do my bidding - p
PixPrucer
09/19/2022, 4:23 PMOh neat the model also learned to do the funny scream - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:24 PM(none of my ideas are in either of those languages) - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:24 PMi am notslippinjimmy (blanka) - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:25 PMyes blankas name is fucking jimmy - c
Couch
09/19/2022, 4:25 PMblanka is jimmy mcgill - u
{K EY1} (Kei)
09/19/2022, 4:25 PMDo you know french Or mandarin - u
(Dawn) Will Draw Fictional Women
09/19/2022, 4:27 PMno